-
Notifications
You must be signed in to change notification settings - Fork 0
/
canny2image_TRT_copy-1659.py
222 lines (192 loc) · 12.5 KB
/
canny2image_TRT_copy-1659.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
from share import *
import config
import cv2
import einops
import gradio as gr
import numpy as np
import torch
import random
import os
import tensorrt as trt
from pytorch_lightning import seed_everything
from annotator.util import resize_image, HWC3
from annotator.canny import CannyDetector
from cldm.model import create_model, load_state_dict
from cldm.ddim_hacked import DDIMSampler
from ldm.util import log_txt_as_img, exists, instantiate_from_config
from cuda import cudart
os.environ['CUDA_MODULE_LOADING'] = 'LAZY'
class hackathon():
def initialize(self):
self.apply_canny = CannyDetector()
self.model = create_model('./models/cldm_v15.yaml').cpu()
self.model.load_state_dict(load_state_dict('/home/player/ControlNet/models/control_sd15_canny.pth', location='cuda'))
self.model = self.model.cuda()
self.ddim_sampler = DDIMSampler(self.model)
H = 256
W = 384
####################################### 配置clip ############################################
# with open("./models/enginemodels/sd_clip_fp32.plan", 'rb') as f:
# clip_engine_str = f.read()
# clip_engine = trt.Runtime(self.trt_logger).deserialize_cuda_engine(clip_engine_str)
# clip_context = clip_engine.create_execution_context()
# clip_tensor_name = [clip_engine.get_tensor_name(i) for i in range(3)]
# # clip的输入
# self.model.cond_stage_model.token=torch.zeros((1,77),dtype=torch.int32).to("cuda")
# clip_context.set_tensor_address(clip_tensor_name[0], self.model.cond_stage_model.token)
# # clip的输出
# self.model.cond_stage_model.last_hidden_state=torch.zeros((1,77,768),dtype=torch.float32).to("cuda")
# clip_context.set_tensor_address(clip_tensor_name[1], self.model.cond_stage_model.last_hidden_state)
# self.model.cond_stage_model.pooler_output=torch.zeros((1,768),dtype=torch.float32).to("cuda")
# clip_context.set_tensor_address(clip_tensor_name[2], self.model.cond_stage_model.pooler_output)
# # CUDA Graph capture
# _, clip_stream = cudart.cudaStreamCreate()
# cudart.cudaStreamBeginCapture(clip_stream, cudart.cudaStreamCaptureMode.cudaStreamCaptureModeGlobal)
# clip_context.execute_async_v3(clip_stream)
# _, clip_graph = cudart.cudaStreamEndCapture(clip_stream)
# _, clip_graphExe = cudart.cudaGraphInstantiate(clip_graph, 0)
# self.model.cond_stage_model.clip_graphExe=clip_graphExe
# self.model.cond_stage_model.clip_graphExe=clip_graphExe
# # in clip do inference with CUDA graph
# cudart.cudaGraphLaunch(self.clip_graphExe, self.clip_stream)
# cudart.cudaStreamSynchronize(self.stream)
################################### 配置controlnet ############################################
self.trt_logger = trt.Logger(trt.Logger.WARNING)
trt.init_libnvinfer_plugins(self.trt_logger, '')
with open("./models/enginemodels/sd_control_fp16.engine", 'rb') as f:
control_engine_str = f.read()
control_engine = trt.Runtime(self.trt_logger).deserialize_cuda_engine(control_engine_str)
control_context = control_engine.create_execution_context()
control_nIO = control_engine.num_io_tensors
control_tensor_name = [control_engine.get_tensor_name(i) for i in range(control_nIO)]
# control的输入
self.model.x_noisy = torch.zeros((1, 4, H//8, W//8),dtype=torch.float32).to("cuda")
self.model.hint_in = torch.zeros((1, 3, H, W),dtype=torch.float32).to("cuda")
self.model.t = torch.zeros((1),dtype=torch.int32).to("cuda") #int32可能有影响
self.model.cond_txt = torch.zeros((1,77,768),dtype=torch.float32).to("cuda")
b, c, h, w = self.model.x_noisy.shape
buffer_device = []
buffer_device.append(self.model.x_noisy.reshape(-1).data_ptr())
buffer_device.append(self.model.hint_in.reshape(-1).data_ptr())
buffer_device.append(self.model.t.reshape(-1).data_ptr())
buffer_device.append(self.model.cond_txt.reshape(-1).data_ptr())
# control的输出
control_out = []
self.model.control_out_1 = torch.zeros(b, 320, h, w, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.control_out_1.reshape(-1).data_ptr())
control_out.append(self.model.control_out_1.reshape(-1).data_ptr())
self.model.control_out_2 = torch.zeros(b, 320, h, w, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.control_out_2.reshape(-1).data_ptr())
control_out.append(self.model.control_out_2.reshape(-1).data_ptr())
self.model.control_out_3 = torch.zeros(b, 320, h, w, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.control_out_3.reshape(-1).data_ptr())
control_out.append(self.model.control_out_3.reshape(-1).data_ptr())
self.model.control_out_4 = torch.zeros(b, 320, h//2, w//2, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.control_out_4.reshape(-1).data_ptr())
control_out.append(self.model.control_out_4.reshape(-1).data_ptr())
self.model.control_out_5 = torch.zeros(b, 640, h//2, w//2, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.control_out_5.reshape(-1).data_ptr())
control_out.append(self.model.control_out_5.reshape(-1).data_ptr())
self.model.control_out_6 = torch.zeros(b, 640, h//2, w//2, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.control_out_6.reshape(-1).data_ptr())
control_out.append(self.model.control_out_6.reshape(-1).data_ptr())
self.model.control_out_7 = torch.zeros(b, 640, h//4, w//4, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.control_out_7.reshape(-1).data_ptr())
control_out.append(self.model.control_out_7.reshape(-1).data_ptr())
self.model.control_out_8 = torch.zeros(b, 1280, h//4, w//4, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.control_out_8.reshape(-1).data_ptr())
control_out.append(self.model.control_out_8.reshape(-1).data_ptr())
self.model.control_out_9 = torch.zeros(b, 1280, h//4, w//4, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.control_out_9.reshape(-1).data_ptr())
control_out.append(self.model.control_out_9.reshape(-1).data_ptr())
self.model.control_out_10 = torch.zeros(b, 1280, h//8, w//8, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.control_out_10.reshape(-1).data_ptr())
control_out.append(self.model.control_out_10.reshape(-1).data_ptr())
self.model.control_out_11 = torch.zeros(b, 1280, h//8, w//8, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.control_out_11.reshape(-1).data_ptr())
control_out.append(self.model.control_out_11.reshape(-1).data_ptr())
self.model.control_out_12 = torch.zeros(b, 1280, h//8, w//8, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.control_out_12.reshape(-1).data_ptr())
control_out.append(self.model.control_out_12.reshape(-1).data_ptr())
self.model.control_out_13 = torch.zeros(b, 1280, h//8, w//8, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.control_out_13.reshape(-1).data_ptr())
control_out.append(self.model.control_out_13.reshape(-1).data_ptr())
for i in range(control_nIO):
control_context.set_tensor_address(control_tensor_name[i], buffer_device[i])
self.model.control_context = control_context
# CUDA Graph capture
# _, control_stream = cudart.cudaStreamCreate()
# cudart.cudaStreamBeginCapture(control_stream, cudart.cudaStreamCaptureMode.cudaStreamCaptureModeGlobal)
# control_context.execute_async_v3(control_stream) #在捕获的时候，异步执行
# cudart.cudaStreamSynchronize(control_stream)
# _, control_graph = cudart.cudaStreamEndCapture(control_stream)
# _, control_graphExe = cudart.cudaGraphInstantiate(control_graph, 0)
# self.model.control_graphExe=control_graphExe
# self.model.control_stream=control_stream
####################################### 配置unet ##############################################
self.trt_logger = trt.Logger(trt.Logger.WARNING)
trt.init_libnvinfer_plugins(self.trt_logger, '')
with open("./models/enginemodels/sd_diffusion_fp16.engine", 'rb') as f:
diffusion_engine_str = f.read()
diffusion_engine = trt.Runtime(self.trt_logger).deserialize_cuda_engine(diffusion_engine_str)
diffusion_context = diffusion_engine.create_execution_context()
diffusion_nIO = diffusion_engine.num_io_tensors
diffusion_tensor_name = [diffusion_engine.get_tensor_name(i) for i in range(diffusion_nIO)]
# diffusion的输入
buffer_device = []
buffer_device.append(self.model.x_noisy.reshape(-1).data_ptr())
buffer_device.append(self.model.t.reshape(-1).data_ptr())
buffer_device.append(self.model.cond_txt.reshape(-1).data_ptr())
for co in control_out:
buffer_device.append(co)
# diffusion的输出
self.model.eps = torch.zeros(1, 4, 32, 48, dtype=torch.float32).to("cuda")
buffer_device.append(self.model.eps.reshape(-1).data_ptr())
for i in range(control_nIO):
diffusion_context.set_tensor_address(diffusion_tensor_name[i], buffer_device[i])
self.model.diffusion_context = diffusion_context
# CUDA Graph capture
# _, diffusion_stream = cudart.cudaStreamCreate()
# cudart.cudaStreamBeginCapture(diffusion_stream, cudart.cudaStreamCaptureMode.cudaStreamCaptureModeGlobal)
# diffusion_context.execute_async_v3(diffusion_stream)
# cudart.cudaStreamSynchronize(control_stream)
# _, diffusion_graph = cudart.cudaStreamEndCapture(diffusion_stream)
# _, diffusion_graphExe = cudart.cudaGraphInstantiate(diffusion_graph, 0)
# self.model.diffusion_graphExe=diffusion_graphExe
# self.model.diffusion_stream=diffusion_stream
######################################### 配置decoder #########################################
print("finished")
def process(self, input_image, prompt, a_prompt, n_prompt, num_samples, image_resolution, ddim_steps, guess_mode, strength, scale, seed, eta, low_threshold, high_threshold):
with torch.no_grad():
img = resize_image(HWC3(input_image), image_resolution)
H, W, C = img.shape
detected_map = self.apply_canny(img, low_threshold, high_threshold)
detected_map = HWC3(detected_map)
control = torch.from_numpy(detected_map.copy()).float().cuda() / 255.0
control = torch.stack([control for _ in range(num_samples)], dim=0)
control = einops.rearrange(control, 'b h w c -> b c h w').clone()
if seed == -1:
seed = random.randint(0, 65535)
seed_everything(seed)
if config.save_memory:
self.model.low_vram_shift(is_diffusing=False)
cond = {"c_concat": [control], "c_crossattn": [self.model.get_learned_conditioning([prompt + ', ' + a_prompt] * num_samples)]}
un_cond = {"c_concat": None if guess_mode else [control], "c_crossattn": [self.model.get_learned_conditioning([n_prompt] * num_samples)]}
shape = (4, H // 8, W // 8)
if config.save_memory:
self.model.low_vram_shift(is_diffusing=True)
self.model.control_scales = [strength * (0.825 ** float(12 - i)) for i in range(13)] if guess_mode else ([strength] * 13) # Magic number. IDK why. Perhaps because 0.825**12<0.01 but 0.826**12>0.01
ddim_steps = 20
samples, intermediates = self.ddim_sampler.sample(ddim_steps, num_samples,
shape, cond, verbose=False, eta=eta,
unconditional_guidance_scale=scale,
unconditional_conditioning=un_cond)
if config.save_memory:
self.model.low_vram_shift(is_diffusing=False)
x_samples = self.model.decode_first_stage(samples)
x_samples = (einops.rearrange(x_samples, 'b c h w -> b h w c') * 127.5 + 127.5).cpu().numpy().clip(0, 255).astype(np.uint8)
results = [x_samples[i] for i in range(num_samples)]
return results
# if __name__=="__main__":
# h = hackathon()
# h.initialize()